A drop in the ocean: Monitoring fish communities in spawning areas using environmental DNA

Early life stages of aquatic organisms are particularly vulnerable to climatic stressors; however, they are difficult to monitor due to challenges in sampling and morphological identification. Environmental DNA (eDNA) from water samples represents an op-portunity for rapid, nondestructive monitoring of aquatic community composition as well as single species monitoring. eDNA can also detect spawning events, although has not been yet tested in offshore spawning grounds . Here, we used metabarcoding of water samples to detect the presence of key fish taxa in spawning areas that are difficult to monitor using traditional means. We analyzed DNA from water samples and fish larvae samples at 14 offshore sites, using 12S mitochondrial metabarcoding and compared taxa detections, diversity, and community structure estimated by both sample types. Species richness and diversity did not differ between water and larvae samples. Both sample types detected a core of 12 taxa across the survey, with an average agreement in detections of 75% at sampling site level. Water samples detected two of the three most abundant taxa, the sandeel, Ammodytes marinus, and clupeids, Clupea harengus/Sprattus sprattus , at 31% and 38% more sites than larvae samples respectively, while Callionymus sp. was more prevalent in larvae samples. Mackerel ( Scomber scombrus ) and blue whiting ( Micromestius poutassou ) were only detected in water samples despite sampling taking place at peak spawning times for these species. Our results demonstrate that eDNA metabarcoding provides a rapid and feasible monitoring method for the management of key taxa, such as sandeel, that cannot be easily monitored using traditional capture surveys.

or limited resources (Costello et al., 2010). In addition, the velocity of climate change and seasonal shifts in timing of temperature changes is, in some cases, greater in the ocean than on land (Burrows et al., 2011); therefore, development of rapid (and feasible) monitoring methods is crucial to detecting the magnitude of these changes (Costello et al., 2010;Thomsen et al., 2012).
Monitoring spawning grounds and fish recruitment is essential in the face of these global pressures as larval and juvenile developmental stages of fishes are highly sensitive to environmental stressors (Pitois, Lynam, Jansen, Halliday, & Edwards, 2012).
Furthermore, information on key taxa that are difficult to capture and identify using traditional trawls (e.g., sandeel) is currently lacking, even in areas that are surveyed more regularly (Ellis et al., 2012;Lynam et al., 2013).
Environmental DNA (eDNA) extracted from water samples is potentially a rapid, cost effective tool for monitoring species distributions (Lodge et al., 2012), thus reducing the need for destructive sampling (Bylemans et al., 2017). Metabarcoding analysis of eDNA can be useful for whole community/broad range assessment (Bohmann et al., 2014;Lacoursière-Roussel et al., 2018;Thomsen et al., 2016) and also perform well for target species monitoring (Harper et al., 2018), offering a potential alternative to traditional monitoring of spawning grounds. The use of eDNA has shown potential for detecting pulses of spawning in freshwater systems where mass release of gametes results in sudden increases of concentration of mitochondrial DNA in the water column (Bylemans et al., 2017). In some cases, increases in eDNA concentration reveal the movement of adults toward a spawning area, rather than the release of gametes (Erickson et al., 2016), but its usefulness to detect spawning in the marine environment remains uncertain.
Although the nondestructive nature of eDNA would make it ideal for use in sensitive environments (Stat et al., 2019), or for the monitoring of rare or threatened species (Bylemans et al., 2017), understanding how eDNA sampling reflects or differs from traditional monitoring techniques can be challenging (Hansen, Bekkevold, Clausen, & Nielsen, 2018). For example, eDNA cannot distinguish among year classes, which in turn can be specifically targeted by physical sampling (Maruyama, Nakamura, Yamanaka, Kondoh, & Minamoto, 2014). In addition, eDNA may show differing dispersal patterns in the open ocean to those physically exhibited by the fish larvae themselves (Goldberg et al., 2016). In fact, comparisons of simultaneous water and visual/physical sampling in the marine environment have found variable level of agreement in detections between eDNA analyses and morphological taxonomy (Leduc et al., 2019;Stat et al., 2019;Thomsen et al., 2016). In order to address these discrepancies, we assessed both eDNA (water samples) and ichthyoplankton physical samples (larvae) using metabarcoding and an identical bioinformatics pipeline. We compared species detection levels, relative abundance, and community composition in both sample types, taken simultaneously at the same locations from the Celtic and Irish seas.

| Field sampling
Water and larvae samples were taken onboard RV Celtic Voyager in the Irish and Celtic seas at 14 sites in known spawning areas for Ammodytidae, Clupeidae, Gadidae, Scombridae, and Pleuronectidae (Ellis et al., 2012) (Figure 1, Table S1). Sampling was carried out in May (17th to 26th, 2018), during or shortly after the spawning season for many of the fish species in the sampling area (Table S2).
Filtration of water samples was carried out in an area physically separated from the processing of larvae samples. Gloves were changed between samples and surfaces cleaned using a 10% bleach solution, before and after each sampling event. Sea surface water samples F I G U R E 1 Map of the sampling locations (14 sites) across the Irish and Celtic sea region. For coordinates of each sampling location, see  Table S1). Ichthyoplankton were separated from other zooplankton species and preserved in RNA later at room temperature for 24 hr, then refrigerated at 4°C until laboratory processing.
Temperature, salinity, and density were recorded at each sampling site (Table S1).

| Laboratory processing-DNA extraction and library preparation
All water samples were extracted using the QIAGEN PowerSoil kit, using a homogenization step (a Precellys 24 tissue homogenizer, Bertin Instruments). Extracted DNA was stored at −80°C until library preparation. Extraction blanks (where no filter was added) were carried through all steps of the library preparation and bioinformatic analysis.
Pooled homogenates of all fish larvae present in the MultiNet haul corresponding to each sampling site were extracted in bulk, by taking a 5 mg (±3 mg) of tissue anterior to the tail (or the complete larva for smaller specimens) from all individuals in a haul (Ratcliffe et al., 2020).

| Bioinformatic analyses
Water, larvae, and mock community and all blanks were subjected to the same bioinformatics pipeline and processed simultaneously. The amplicon sequence variant (ASV) approach was used because it enables detection of single nucleotide differences (Callahan, McMurdie, & Holmes, 2017) and therefore provides a higher resolution than a traditional OTU approach.
Qiime2 (version 2019.1, Bolyen et al., 2019) was used to process de-multiplexed paired end sequences. DADA2 (Callahan et al., 2016) within Qiime2 was used for de-noising steps. Based on read quality scores, the first 10 bp of each sequence was trimmed and all sequences truncated to 100 bp in length. Subsequently, sequencing errors were corrected where possible, and chimeric sequences removed, paired end reads were joined and sequences de-replicated using the default DADA2 settings in Qiime2.
Taxonomic assignment was conducted using custom databases (Ratcliffe et al., 2020). Initially, reads were classified against a full database that included all taxa available on NCBI amplified in silico with the 12S V5 primers, using the KNN method in mothur (Schloss et al., 2009). A second screening was then carried out using a smaller database that included all available sequences of fish encountered (native and non-native) in the British Isles (Fish Base: accessed 31/3/2019) as well as outgroups known to be present from the classification against the full database, again using the KNN method, and the parameter "numwanted = 1" (Findley et al., 2013). Due to the potential for false positive assignments using this parameter, these assignments were verified using NCBI megablast (Morgulis et al., 2008), where the top 10 hits were screened on a case by case basis, for the highest match of a UK fish taxon. Assignments below 98% identity were assigned to genus or family level using the lowest common ancestor algorithm (Huson, Auch, Qi, & Schuster, 2007) in MEGAN (6.15.1).
Once taxonomic assignment was complete, non-fish ASVs were removed from downstream analysis. To remove contaminant ASVs (false positives) from the data, filtration blank read counts (subjected to the same workflow at all steps of the process as field samples) were subtracted from each field replicate in the corresponding site, . This was carried out before any downstream data analysis (Andruszkiewicz et al., 2017).

| Statistical analysis
All statistical analysis was performed using R (version 3.5.3, R Core Team, 2017). For comparison purposes, as only one larvae haul was sampled per site, the reads from the three water biological replicates and three larvae technical replicates were pooled for each site. Reads were then converted to relative abundance data (proportion) to account for unequal sequencing depths between samples. Subsequently, to remove any reads present due to tag jumping (Schnell, Bohmann, & Gilbert, 2015) taxa that accounted for <0.05% (set using the tag jumping blank) of the relative abundance of a sample were removed from that site for downstream analysis.
Mock communities were analyzed using a chi-square (goodness of fit) test to ascertain if there was any difference between relative abundance of genomic DNA in the sample (expected) and relative abundance of reads after sequence processing (observed). Species richness was calculated for each sample using "specnumber" in R, and Shannon Wiener diversity was calculated using the "diversity" function with "method" = "Shannon." Wilcoxon signed rank tests (paired samples) were used to ascertain if medians differed between water and larvae samples and variances within each treatment were calculated as median absolute deviation (MAD), using r function "mad." After checking the residuals for normality using "skewness" (Moments package), an F test was used to check for significant differences in variance between the sample types. Log-likelihood ratio (G-test) test of independence with Williams' correction was used to test the effect of sample site and taxon on detections between the two methods.
To assess differences in community composition between the two methods, a dummy number of species of 1 was added to all samples in order compute Brae-Curtis dissimilarity for sites where no larvae were captured (Clarke, Somerfield, & Chapman, 2006). Reads (relative abundance) were square-root transformed, a Bray Curtis dissimilarity matrix was generated and PERMDISP (to test for homogeneity of dispersion) and subsequent PERMANOVA analysis was used (Anderson, 2014) ( Figure S1). SIMPER analysis was then carried out on untransformed relative abundance data to ascertain which species were driving the differences observed between the two sampling methods (Clarke, 1993).  Total number of reads (by colour)
In both water and larvae samples, we successfully detected 12 of the same taxa ( Figure 2) and a similar number of taxa (19 taxa in water, 20 taxa in larvae) were detected overall. At a site level, there was an average of 75% agreement in taxa detection between the sample types (Table 2; Figure 3). The number of taxa detected in the two sample types depended on the sampling site (G = 31.43, df = 12, p = .002) and the taxon considered (G = 42.80, df = 27, p = .027, Figure 3, Table 2). In general, the more abundant taxa were detected by both sampling methods while less abundant taxa exhibited much greater variance, with some taxa being detected in one sample type.
Of these, 10 taxa were detected at only one site, 60% of which were observed in larvae samples ( Figure 3, Table 2).
Patterns of relative abundance broadly followed those observed in the number of detections. C. harengus/S. sprattus, A. marinus were detected in higher relative abundance in water than in larvae samples (Figure 2 Table S5). Taxon richness and alpha diversity did not differ between the two sampling methods (

| D ISCUSS I ON
We applied water eDNA metabarcoding to detect fish in marine spawning areas and demonstrated that water samples not only broadly reflect larvae samples, with 75% average agreement on site by site basis but can also be more sensitive in the case of particular taxa, such as sandeels and herring/sprat. While rare species, those detected at one site only, were more likely to be found in bulk samples, some taxa expected to be part of the spawning assemblage (e.g., mackerel S. scombrus and blue whiting, Micromestius poutassou) were only detected in water samples, highlighting the potential of this tool to complement traditional sampling.
Previous studies that compare water sampling to visual or capture surveys have found varying levels of agreement between the taxa detected (Cilleros et al., 2019;Stat et al., 2019;Thomsen et al., 2016). Thomsen et al., (2016) found correlations between capture biomass and numbers of water sample reads, while Stat et al., (2019) found fish assemblages differed between visual (BRUV) and water samples.
Differences in detection between traditional methods and eDNA are influenced by eDNA dispersal range, year class, and detection sensitivity of the particular methods compared. Importantly, all monitoring techniques are subject to biases, for example trawl types may also differ in the species captured due to gear selectivity biases (Hansen et al., 2018). Hence, while water samples may not always perfectly reflect capture/visual samples, they represent a rapid and feasible and nondestructive way to efficiently assess fish community assemblages (Cilleros et al., 2019;Thomsen et al., 2016).
Here, we compared surface water samples to larvae sampled between the surface and 50 m depth. Fish eggs and larvae are most abundant in depths shallower than 50 m (Conway, Coombs, & Smith, 1997;Sabatés, 2004). In the Irish Sea, densities of fish eggs increase with decreasing depth and peak at the surface, while larval density increases toward a peak at of 10-15 m, with little difference observed between species, including the families most abundant in this study (Clupeidae, Ammodytidae, Callionymidae) (Conway et al., 1997). For some taxa, however, this general pattern may not apply; for instance, hake larvae (M. merluccius) have been shown aggregate at maximum density at depths of 60-80 m (Sabatés, 2004). This taxon was not detected in water samples, despite being detected in one larvae haul; therefore, while sampling depth is unlikely to have affected the majority of taxa, it may be a reason for some of the discrepancies in the detection of rare species. Thus, as eDNA exhibits sensitivity to vertical zonation in stratified water (Jeunen et al., 2020), multiple sampling depths may be advisable, depending on the taxa and life stages of interest.
For sensitive taxa such as sandeels that are hard to monitor using traditional means (Ellis et al., 2012), this survey demonstrates the potential of eDNA metabarcoding as a monitoring tool.
Sandeels are a key prey species, consumed by fish, seabirds, and marine mammals; however, due to their short life cycle, and the reliance of their stocks on larval recruitment, these taxa are difficult to sample and the stocks are therefore difficult to manage (Lynam et al., 2013). Neither otter nor beam trawls are effective methods for assessing their abundance, particularly at early life stages (Ellis et al., 2012). In addition, morphological identification is often unreliable (Thiel & Knebelsberger, 2016), limiting the assessment to family level only (Ellis et al., 2012). Our survey encountered the lesser sandeel, A. marinus, a taxon of the genus Gymnammodytes and a further taxon identified to family level, distributed in areas where sandeels are known to spawn in the Irish Sea and the Bristol channel (Ellis et al., 2012;Lynam et al., 2013). A. marinus was always detected in water samples where the larvae were also encountered. Therefore, while water sample metabarcoding alone cannot determine the age class of organisms encountered, it can give a picture of the distribution of these taxa during spawning events, to a higher taxonomic resolution than is often available through traditional means (Ellis et al., 2012).
Herring/sprat (C. harengus/S. sprattus) were also frequently encountered using both sample types at the same sites, with water samples displaying higher sensitivity. Most C. harengus spawning in the Irish Sea occurs in September/October, and some can spawn as late as March (Brophy & Danilowicz, 2002). However, due to difficulties in morphologically separating these two species, spring surveys tend to assume that larvae caught in this period are S. sprattus (Fox, Dickey-Collas, & Winpenny, 1997), which spawn from March to August (de Silva, 1973). While the primers used in this study cannot separate S. sprattus and C. harengus, a qPCR approach could be used in the samples where the presence of one or both species is identified by metabarcoding, to rapidly separate these species in water samples (e.g., Brechon et al., 2013). S. sprattus is relatively under-assessed within the Celtic Seas ecoregion and is considered data limited (Moore et al., 2019); therefore, information obtained from a water sampling approach has the potential to add to traditional methods of assessment.
Relative abundances were correlated for L. limanda between the two sample types. However, relative abundances estimated from metabarcoding should be treated with caution (Lamb et al., 2019;Thomas, Deagle, Eveson, Harsch, & Trites, 2016) due to amplification bias. In fact, the sequencing of mock community 1 indicated that, with the primers used in this study, the relative abundance of A. marinus reads could be on average 2 times higher than the relative abundance of input material, while Lepidorhombus sp. reads were 0.6 times as abundant as their known inputs. In addition, DNA shedding may differ between organisms (Sassoubre, Yamahara, Gardner, Block, & Boehm, 2016) and sizes of fish (Maruyama et al., 2014), and can further complicate abundance signals (Hansen et al., 2018).
However, using a presence/absence approach can also overestimate the importance of rare taxa and relative abundance estimates may provide useful information (Deagle et al., 2019;Lamb et al., 2019) for comparative studies. In this case, C. harengus/S. sprattus and A. marinus were more frequently detected, and had a higher relative abundance, in water samples, while Callionymus sp. was more frequently detected in bulk samples, but no difference in relative abundance was observed between sample types. This indicates that, in some contexts, measures of relative abundance can provide useful information.
For rarer species, detections by both methods were more sporadic. For example, taxa such as mackerel and blue whiting, which are known to undergo peak spawning in May (Ellis et al., 2012), were not encountered in larvae samples, but were detected in water samples. While it is not possible to know whether spawning had occurred and was missed by the larvae sampling, or whether only adults were present, this indicates that taxa potentially missed using traditional means can still be detected in water samples.
eDNA metabarcoding detections are sensitive to computational filtering thresholds (Evans et al., 2017). In this case, blank filtering  and discarding of taxa with low reads to account for index hopping (Schnell et al., 2015) resulted in less water than larvae sample detections for some of the less common/abundant taxa for example M. merluccius, and L. limanda (Figure 3). This reflects the trade-off between stringency and uncertainty when applying thresholds of detection to metabacoding data, as while more stringent filtering can underestimate taxa richness, it also reduces the risk of false positives due to tag jumping (Schnell et al., 2015) or cross contamination. Potential solutions include using combinations of markers (Evans et al., 2017) or specific qPCR assays (Harper et al., 2018;Schneider et al., 2016) in conjunction with metabarcoding .
While water and larvae samples did not differ overall in richness and alpha diversity measures, larvae samples exhibited greater variance in species richness, community composition, and detection of rare taxa differed between sample types, demonstrating how these two sampling strategies may complement each other. When considering how to interpret eDNA data, the ecology of the eDNA molecules should be considered (Barnes & Turner, 2016). eDNA transport in offshore areas has not be studied extensively (Collins et al., 2018); however, in freshwater systems, eDNA signals may travel up to 10 km (Deiner & Altermatt, 2014). In the marine environment, tidal currents, seasonal stratification, pH, and temperature (Lacoursière-Roussel et al., 2018) may all influence the distribution of eDNA in the marine environment (Hansen et al., 2018). In coastal environments, with a small tidal amplitude, eDNA has been shown to have a limited dispersion area, <1,000 m, and may only remain detectable for as little as an hour after the source has been removed, providing a snapshot of the organisms present (Murakami et al., 2019). In offshore environments however, eDNA may degrade more slowly than in coastal areas (Collins et al., 2018). Environmental factors can, therefore, lead to widely variable dispersal of eDNA particles, dependent on oceanographic, biological and chemical parameters (Hansen et al., 2018).

| CON CLUS IONS
Water and larval sampling both coincided in the detection of the most common taxa, which constituted 63% and 60% of the taxa in the water and larvae samples, respectively. On a site by site basis, there was a 75% agreement in detection between sample types.
Sandeels were detected more frequently in water samples, which improves upon traditional methods that often are unable to capture or identify this family. Herring/sprat were also more frequently detected in water samples, indicating that metabarcoding combined with a targeted approach such as qPCR could also provide higher sensitivity distributions for these taxa. While eDNA still requires an extensive sampling effort, its noninvasive and rapid nature renders it particularly suitable for use in spawning and protected areas and for fisheries management applications.

ACK N OWLED G M ENTS
We are very grateful to those who assisted with sampling: Helen McCormick, Ross O'Neill, Michael Sheridan, Sarah Albuixech-Marti, Katie Costello, and crew of the R.V. Celtic Voyager. We also thank Deiene Rodriguez-Barreto and Jessica Minett for assisting with the water sample methodology and Richard O'Rorke for making the custom taxonomy database.

CO N FLI C T O F I NTE R E S T
None declared.